anonymousPaper5674/Property-Driven-GNN

一个用于评估图神经网络（GNNs）在16个基本图属性上表达能力的大规模基准数据集。该数据集包含352个图分类数据集，分为两个家族：GraphRandom（176个数据集），其中的图要么满足给定属性，要么随机违反；GraphPerturb（176个数据集），每个正样本图都与一个结构相似但仅差一两条边的负样本图配对。所有数据集均使用Alloy（一种关系逻辑规范语言和基于SAT的分析器）生成，确保了正样本的穷举枚举和负样本的可验证性。数据集格式为CSV文件，每行编码一个图，包括邻接矩阵的扁平化表示和二进制标签（1表示满足属性，0表示不满足）。图是有向的，允许自环。A large-scale benchmark for evaluating the expressive power of Graph Neural Networks (GNNs) across 16 fundamental graph properties. The benchmark contains 352 graph-classification datasets organized into two families: GraphRandom (176 datasets) with graphs that either satisfy or randomly violate a given property, and GraphPerturb (176 datasets) where each positive graph is paired with a structurally similar negative counterpart that differs by only one or two edges. All datasets were generated using Alloy, a relational-logic specification language and SAT-based analyzer, ensuring exhaustive enumeration of positive samples and verifiable negative samples. Data is in CSV format, each row encoding one graph with flattened adjacency matrix and binary label (1 for property satisfaction, 0 otherwise). Graphs are directed and allow self-loops.